Normalized translation

[wanyiming2017]

Hi, in the paper, you wrote "We resolve the scale ambiguity by normalizing translations and depth values such that ||t||=1."  Now, I want to obtain actual translation (not normalized). Is it possible?

To obtain the real depth map , I invert the pixel value of the output depth map, multiply it with the scaling factor s. However, the result is not very good. I wonder if I have made some mistakes here?

[benjaminum]

The returned depth map has already been scaled with the scaling factor.

It is not possible to obtain a depth map with a real world unit. In general this is not possible. E.g. you can never know if an image shows a house or a small toy house.

[wanyiming2017]

Thanks for your reply. So neither can I obtain the translation with a real world unit? If I want to use the estimated translation and depth map to reconstruct 3D scene， the translation and depth map should be in the same scale. I directly use the following three estimated value, is it right? rotation = result['predict_rotation'] translation = result['predict_translation'] depth_map=1/result['predict_depth0']

[Xinrui-Fang]

I also have a quesiton about this issue, is there any possibe solution to get the translation in real world? As the Figure 7 in the paper, author showed the camera position, is there a possibe way to solve this issue?